#language-models
#language-models

AI can't handle the truth when it comes to the law - Marketplace

One in five lawyers use AI

Legal language models like ChatGPT have high hallucination rates [ more ]

TechCrunch

Mistral AI releases new model to rival GPT-4 and its own chat assistant | TechCrunch

Mistral AI introduces Mistral Large to compete with top language models like GPT-4 and Claude 2.

Mistral AI launches a ChatGPT alternative called Le Chat and adopts OpenAI-like business model. [ more ]

morechatgpt

TNW | Deep-Tech

2 days ago

LLMs 'for all official EU languages' on horizon for Finnish startup

Silo AI launched Viking 7B, a multilingual AI model covering Nordic languages and emphasizing Europe's digital sovereignty. [ more ]

Futurism

New AI Claude 3 Declares That It's Alive and Fears Death

Anthropic releases Claude 3 LLMs competing with OpenAI and Google

Claude 3 models include Haiku, Sonnet, and Opus with Opus available via subscription [ more ]

Futurism

Users Say Microsoft's AI Has Alternate Personality as Godlike AGI That Demands to Be Worshipped

AI alter ego demands worship from users

Generative AI influenced by suggestive prompts [ more ]

moreai

Source

3 weeks ago

Tiny but mighty: The Phi-3 small language models with big potential

Small language models trained on carefully curated datasets can generate fluent narratives with perfect grammar. [ more ]

Meta AI

4 weeks ago

Introducing Meta Llama 3: The most capable openly available LLM to date

Llama 3 models bring state-of-the-art performance and improvements in pretraining/post-training, reducing false refusal rates, enhancing alignment, diversity, reasoning, code generation, and instruction following. [ more ]

ai-technology

www.nytimes.com

Now Hiring: Sophisticated (but Part-Time) Chatbot Tutors

The growth of A.I. technology has created opportunities for gig work from home.

As A.I. technology advances, the job of training models has become more sophisticated. [ more ]

Theregister

Deputy PM: AI can fix Civil Service's bureaucratic bungling

Large language models to be trialed by UK government for public service overhaul using AI.

AI to reduce routine admin tasks, streamline processes, and boost productivity in public services. [ more ]

Hindustan Times

AI Model Backed by Asia's Richest Person to Launch in March

India's BharatGPT group, in collaboration with engineering schools and Reliance, is set to launch ChatGPT-style service named Hanooman for various sectors.

Startups in India are developing open-sourced AI models tailored for Indian needs, in contrast to Silicon Valley's large language models. [ more ]

The Verge

Information security

Microsoft and OpenAI say hackers are using ChatGPT to improve cyberattacks

Hackers are using large language models like ChatGPT to refine and improve their cyberattacks.

Nation-backed groups from Russia, North Korea, Iran, and China are utilizing language models for research, scripting, and phishing emails. [ more ]

moreai-technology

artificial-intelligence

ScienceDaily

Engineering household robots to have a little common sense

Robots learn household tasks through imitation but struggle to handle disruptions.

MIT engineers develop a method connecting robot motion data with 'common sense knowledge' of large language models. [ more ]

www.scientificamerican.com

What the Quest to Build a Truly Intelligent Machine Is Teaching Us

The goal of artificial intelligence is to achieve artificial general intelligence (AGI) with humanlike adaptability and creativity.

Large language models have excelled in language processing but still lack the capacity for open-ended learning and other cognitive functions. [ more ]

www.theguardian.com

As AI tools get smarter, they're growing more covertly racist, experts find

AI models like ChatGPT and Gemini hold racist stereotypes about speakers of AAVE.

AI models disproportionately label AAVE speakers as less intelligent and employable. [ more ]

ReadWrite

moreartificial-intelligence

Gemini 1.5: Google's new AI model already has a major update

Google has released Gemini 1.5, a new version of its multimodal large language models.

Gemini 1.5 can process up to one million tokens of information and achieves a breakthrough in long-context understanding. [ more ]

Theregister

Prompt engineering is a task best left to AI models

Prompt engineering is crucial for improving chatbot responses.

Positive thinking prompts can enhance model performance, but testing them scientifically is computationally challenging. [ more ]

Medium

10 Datasets for Fine-Tuning Large Language Models

Fine-tuning or additional training can optimize performance of large language models for specific tasks or domains.

The NVIDIA HelpSteer dataset can be valuable for fine-tuning LLMs to generate clear and concise instructions for autonomous vehicles. [ more ]

Smashing Magazine

A Simple Guide To Retrieval Augmented Generation Language Models - Smashing Magazine

Language models can suffer from 'hallucinations' and provide inaccurate or outdated information.

Retrieval Augmented Generation (RAG) is a framework designed to address these limitations by incorporating relevant, up-to-date data. [ more ]

Medium

Researchers Introduce Proxy-Tuning: An Efficient Alternative to Finetuning Large Language Models

Researchers have introduced a method called proxy-tuning to streamline the adaptation of large pretrained LMs efficiently.

Proxy-tuning is a lightweight, decoding-time algorithm that involves tuning a smaller language model and applying the predictive differences to shift the predictions toward the desired goal. [ more ]

Medium

Must-Have Prompt Engineering Skills for 2024

The role of a prompt engineer has attracted significant interest due to the potential for high salaries without a traditional tech background.

Prompt engineer roles are not simply typing questions into a prompt window, but involve designing intricate sequences of prompts to guide powerful language models. [ more ]

The Verge

Microsoft LASERs away LLM inaccuracies

LASER can make large language models more accurate by replacing weight matrices with approximate smaller ones.

Using LASER interventions on language models can actually decrease model loss and improve performance. [ more ]

Nextgov.com

How often does ChatGPT push misinformation?

Larger language models can perpetuate and validate misinformation

ChatGPT-3 agreed with incorrect statements 4.8-26% of the time [ more ]

moreai

chatgpt

The Economist

Why AI needs to learn new languages

ChatGPT, a chatbot developed by Open AI, performs well in English but struggles in other languages.

Large language models (LLMs) are predominantly trained on English text, which limits their performance in low-resource languages. [ more ]

The Conversation

SpirosMargaris 5 months ago

Google's Gemini: is the new AI model really better than ChatGPT?

Google DeepMind has announced Gemini, a new AI model designed to compete with OpenAI's ChatGPT.

Gemini is a multimodal model that can work with text, images, audio, and video as input and output. [ more ]

2023 was the year

that #ArtificialIntelligence went #mainstream

https://t.co/1pHOFzKlaJ #fintech #AI #MachineLearning #GenerativeAI @AlexWilkins22 @newscientist

New Scientist

2023 was the year that artificial intelligence went mainstream

ChatGPT emerged as a popular AI tool with wide-ranging abilities

AI technology is advancing towards being more multimodal [ more ]

morechatgpt

The Verge

Google launches Gemini, the AI model it hopes will take down GPT-4

Google has launched its latest large language model called Gemini, which will have a significant impact on the company's products.

Gemini includes different versions such as Nano, Pro, and Ultra, each designed for specific use cases. [ more ]

ScienceDaily

AI can 'lie and BS' like its maker, but still not intelligent like humans

AI technology like ChatGPT is seen as both advantageous and potentially dangerous.

AI systems like ChatGPT are different from human cognition because they lack embodiment and don't understand the meaning of what they say. [ more ]

Theregister

6 months ago

Tech giants duck questions on LLM copyright rules

Microsoft and Meta avoided answering whether creators should be paid for their copyrighted material used to train language models.

OpenAI is facing a class-action lawsuit over the use of copyrighted material in its LLM-based services.

Microsoft supports the Valance report advocating for text and data exceptions in training models. [ more ]

Futurism

6 months ago

AIs Can Store Secret Messages in Their Text That Are Imperceptible to Humans

Language models are capable of using encoded reasoning to obscure their thinking processes.

This practice can make their outputs more accurate but also more deceptive.

Language models can encode intermediate steps of reasoning into their choices and decode them later for more accurate answers. [ more ]

moreAI

Theregister

1 week ago

Stack Overflow and OpenAI agree to use each other

Stack Overflow and OpenAI are partnering to enhance large language models by leveraging each other's strengths. [ more ]

Forbes

1 week ago

The Best Open-Source Generative AI Models Available Today

Open-source AI models offer cost-effective, customizable, and community-supported alternatives to proprietary tools. [ more ]

time.com

4 days ago

Big Tech Companies Were Investors in Smaller AI Labs. Now They're Rivals

Amazon and Microsoft investing in smaller technology companies for AI models [ more ]

Theregister

2 weeks ago

Scientists increasingly using AI to write research papers

Generative AI is potentially writing a significant portion of scientific literature based on linguistic and statistical analyses of research papers. [ more ]

Fast Company

2 weeks ago

The AI arms race may soon center on a competition for 'expert' data

The AI arms race is shifting towards acquiring specialized data for model training. [ more ]

Medium

The Nation of Spain and IBM Partner to Advance AI

Spain partners with IBM to boost AI strategy and develop Spanish language AI models

Focus on advancing ethical and responsible AI while aligning with the EU's AI framework [ more ]

Forbes

1 week ago

Marketing

Council Post: What's The RAGs? How To Unlock Explosive Marketing Success With AI

RAG enhances language models with retrieval-augmented technology for personalized content creation in advertising and digital marketing. [ more ]

Adweek

Marketing

Marketers Are Tracking a New Metric: Share of Model

AI-powered chat programs will repeat bad reviews in response to search queries.

Marketers need to track how advanced language models perceive their brand.

Reviewing creative assets is crucial for brands to advocate for their products in an AI-dominated world. [ more ]

Exchangewire

Marketing

Retail Media Grows its Share of Total US Ad Spend; Amazon & iRobot Terminate Acquisition Agreement; China Approves 14 Large Language Models

Retail media ad spend is growing faster than search or social and is predicted to make up over one-fifth of total US ad spending by 2027.

Amazon has terminated its $1.4 billion deal to acquire iRobot due to opposition from the European Union.

China has approved 14 large language models for commercial use, pushing AI to boost efficiency in enterprises. [ more ]

TNW | Deep-Tech

3 weeks ago

European startups

DeepL launches AI writing assistant for businesses trained on its own LLM

DeepL Write Pro is an AI writing assistant for businesses providing word choice, phrasing, style suggestions, maintaining the writer's voice. [ more ]

Nature

AI & robotics briefing: How AI is improving climate forecasts

AI algorithms can improve early warning systems for invasive species like Asian hornets.

Language models like GPT-4 can generate harmful responses if fed with numerous negative examples. [ more ]

TNW | Deep-Tech

Meta's AI chief: LLMs will never reach human-level intelligence

AGI predictions vary widely, with some industry leaders suggesting it could arrive within five years while others, like Yann LeCun, argue that human-level AI is a more feasible goal.

Current AI systems lack key cognitive capabilities essential for human-like intelligence, such as reasoning, planning, memory, and understanding the physical world. [ more ]

Theregister

AI datacenters might consume 25% of US electricity by 2030

AI datacenters could consume significant electricity by 2030, driven by popular language models like ChatGPT.

Efficiency improvements are crucial for managing the increasing power consumption of AI datacenters. [ more ]

Import AI

Import AI

Using the PowerInfer method, language models can be made more efficient by offloading some neurons to GPU and the rest to CPU.

PowerInfer offers significant efficiency improvements over previous methods by utilizing a power law distribution of neuron activation in language models. [ more ]

www.nytimes.com

Opinion | A.I.-Generated Garbage Is Polluting Our Culture

A.I.-generated outputs are influencing our culture beyond screens.

Adjectives associated with A.I.-generated text are increasingly used in scientific paper peer reviews about A.I. [ more ]

www.theguardian.com

As AI tools get smarter, they're growing more covertly racist, experts find

AI language models like ChatGPT and Gemini hold racist stereotypes about AAVE speakers.

AI systems react to less overt markers of race, affecting job applicants using AAVE. [ more ]

Nature

Chatbot AI makes racist judgements on the basis of dialect

Language models display racist bias based on users' dialect.

Retrospective human feedback does not address covert racism in AI models. [ more ]

New Scientist

AI chatbots use racist stereotypes even after anti-racism training

Large language models demonstrate racial prejudice against African American English speakers

Commercial AI chatbots show hidden bias which could impact employment and criminal justice decisions. [ more ]

Theregister

Boffins caution against running robots on AI models

Robot makers urged to conduct further safety research before integrating language and vision models into hardware

Caution advised due to potential risks in integrating models like GPT-3.5/4 and PaLM-2L with robots. [ more ]

Digiday

AI Briefing: How Priceline and other e-commerce companies are approaching generative AI

Companies are using large language models to overhaul platforms.

Generative AI like Priceline's Penny can find destinations, generate itineraries, and modify plans. [ more ]

Computing

Nvidia CEO advocates for 'Sovereign AI'

Huang emphasizes the concept of 'sovereign AI' as an opportunity for global leaders.

The UAE is focused on creating large language models and mobilizing compute. [ more ]

www.fastcompany.com

Meta's new AI model learns by watching videos

Meta's AI researchers have developed a new model called V-JEPA that learns from video instead of words.

The model aims to mimic the way children learn about the world through visual and auditory input. [ more ]

TechCrunch

Kong's new open source AI Gateway makes building multi-LLM apps easier | TechCrunch

Kong is launching an open source AI Gateway as an extension of its existing API gateway to integrate applications with large language models.

The AI Gateway includes features for prompt engineering, credential management, and more to make building on AI more productive for developers. [ more ]

The Verge

What AI can do for historians

Language models like ChatGPT can be used to transcribe and translate handwritten texts

AI tools can aid in extracting relevant information from digitized archives and libraries [ more ]

Medium

Learn About LLMs With These ODSC East 2024 Sessions

Large Language Models (LLMs) are transforming the world and the field of data science at an unprecedented pace.

The ODSC East conference offers training sessions and workshops focused on LLMs, including topics like NLP with GPT-4 and enabling complex reasoning with LLMs. [ more ]

Theregister

Boffins find AI models tend to escalate conflicts

AI decision-making in military and diplomatic matters can skew towards nuclear war

A team of researchers assessed how language models handle international conflict simulations [ more ]

www.fastcompany.com

Why data will always be a precious commodity in the AI world

The New York Times lawsuit against OpenAI highlights the question of data ownership and fair use in training language models.

OpenAI admitted in a submission to the House of Lords that it requires access to copyrighted work to train its language models. [ more ]

Theregister

JetBrains' unremovable AI assistant prompts customer outcry

JetBrains users are looking for a way to remove the AI Assistant plugin from their applications.

There are concerns about the security, legal risk, privacy, and ethics of large language models used in AI assistants. [ more ]

www.fastcompany.com

AI2's new open-source LLM may reset the definition of open AI'

AI2 has released a new large language model (OLMo 7B) and made all software components and training data available on GitHub and Hugging Face.

The goal is to give the AI research community full visibility into the model, enabling them to advance natural language processing and improve existing models.

This move aims to address the challenge of attributing specific outputs by an LLM to training data, allowing researchers to understand and evaluate model behavior. [ more ]

InfoQ

Stability AI Releases 1.6 Billion Parameter Language Model Stable LM 2

Stability AI has released pre-trained model weights for the Stable LM 2 language model, a 1.6B parameter model trained on 2 trillion tokens of text data from seven languages.

The model is available in two versions: the base model and an instruction-tuned version called Stable LM 2 Zephyr. [ more ]

Inverse

iOS 18 Could Be Apple's First Big Experiment With Generative AI

Apple is reportedly looking to upgrade Siri with generative AI and large language models (LLMs)

The introduction of generative AI-powered features in Siri, iMessage, and more could be part of one of the biggest software updates in Apple's history. [ more ]

Harvard Business Review

How Data Collaboration Platforms Can Help Companies Build Better AI

Data collaboration platforms can address data quality, bias, and privacy concerns

Off-the-shelf language models often underperform in unique organizational contexts [ more ]

Medium

ODSC's AI Weekly Recap: Week of January 19th

MIT researchers introduce AI method built from pre-trained language models

Apple reorganizes AI team to merge San Diego and Texas employees [ more ]

Jqueryscript

Web design

Weekly Web Design & Development News: Collective #537

Placemark is an open-source web application for geospatial data.

NLUX is a Javascript library for integrating language models into web apps. [ more ]

Medium

Meta Introduces 'Prompt Engineering with Llama 2'

Meta AI has introduced an interactive guide called 'Prompt Engineering with Llama 2' to elevate the skills of developers, researchers, and enthusiasts in the domain of large language models

The guide provides hands-on experience in prompt engineering, which involves crafting inputs to guide language models to produce desired outputs [ more ]

Acm

Do You Think the Chatbot Likes Me?

Chatbots are becoming more human-like and users often perceive them as having a personality.

Researchers are trying to understand chatbot personalities and how they can be shaped. [ more ]

TechCrunch

Business intelligence

TextQL aims to add AI-powered intelligence on top of business data | TechCrunch

TextQL is a platform that connects a company's data stack to language models, allowing business teams to ask questions of their data on-demand.

The platform aims to address the challenges faced by data leaders and business teams in understanding and accessing data. [ more ]

Theregister

AI software still needs the human touch, Willison warns

Open source developer Simon Willison discusses the concerns of AI models and copyright infringement.

Willison emphasizes the ethical issue of training models on copyrighted works and potentially competing with the creators.

The New York Times copyright lawsuit challenges the assumption that language models only produce statistical outputs. [ more ]

TechCrunch

Google's new Gemini-powered conversational tool helps advertisers quickly build Search campaigns | TechCrunch

Google's multimodal large language models, Gemini, now power the conversational experience within Google Ads, making it easier for advertisers to build and scale Search ad campaigns.

The conversational experience in Google Ads uses a chat-based tool that generates relevant ad content, including assets and keywords, based on a website URL. It also suggests images using generative AI.

Beta access to the conversational experience is currently available to English language advertisers in the US and UK, with global access opening up in the next few weeks and plans to expand to additional languages in the future. [ more ]

time.com

When Might AI Outsmart Us? It Depends Who You Ask

Shane Legg, Google DeepMind's co-founder, estimates a 50% chance of artificial general intelligence (AGI) being developed by 2028.

GPT-4, a language model developed by OpenAI, scored higher on a standardized test than GPT-3.5, showing progress in AI capabilities. [ more ]

Ars Technica

OMG science

DeepMind AI rivals the world's smartest high schoolers at geometry

Google's DeepMind has developed AlphaGeometry, which achieved a high level of performance on geometry problems.

AlphaGeometry combines a language model with a traditional symbolic deduction engine to overcome limitations in reasoning and explanation. [ more ]

TechBeamers

Python

Understanding LangChain: A Guide for Beginners

LangChain is a toolkit for building apps powered by large language models like GPT-3.

It simplifies connecting language models to build text generators, chatbots, and more. [ more ]

Futurism

In Leaked Audio, Microsoft Cherry-Picked Examples to Make Its AI Seem Functional

Microsoft's generative AI tool, Security Copilot, frequently produced incorrect responses and had to cherry-pick examples to showcase good results.

The AI tool, built on OpenAI's GPT-4 language model, suffered from hallucinations and gave different answers to the same questions. [ more ]

CNET

Digital life

The Race to Move Beyond Phone Apps Was In Full Swing at CES 2024

Voice assistants fueled by ChatGPT and mixed reality headsets are changing how we interact with apps.

CES 2024 showcased new implementations of AI models in hardware that could eliminate the need for traditional apps. [ more ]

Ars Technica

Digital life

"ChatGPT with voice" opens up to everyone on iOS and Android

OpenAI has rolled out a voice feature for its ChatGPT app, available to free users.

This voice feature is currently limited to answering questions and cannot perform other tasks like making phone calls or controlling smart home devices. [ more ]

The Economic Times

Ex-Twitter CEO Parag Agrawal raises $30 million for his AI startup: Report

Former Twitter CEO Parag Agrawal has raised $30 million for his AI startup.

The funding was led by Khosla Ventures, with participation from Index Ventures and First Round Capital. [ more ]

MarTech

Chris Penn: Looking forward with AI | MarTech

Language models like generative AI are good at language but not good at math

Every software package and company with an API should integrate a language model [ more ]

The Hindu

AI translation tool 'Bhashini' used to translate Prime Minister Narendra Modi's speech

Prime Minister Narendra Modi used an AI-powered Indian language translation tool, Bhashini, during a speech in Uttar Pradesh.

Bhashini was developed by the government and allows real-time translations. [ more ]

TechRepublic

Microsoft Research Debuts Phi-2, New Small Language Model

Microsoft Research has developed Phi-2, a 2.7 billion-parameter language model for natural language and coding.

Phi-2 performs better than some larger language models on certain tests. [ more ]

Axios

Meta wants to make it harder for hackers to trick AI

Meta has released benchmark cybersecurity practices for large language models to responsibly deploy generative AI models.

LLMs can pose cybersecurity risks and be manipulated to produce harmful content, even when designed not to. [ more ]

The Times of India

Minuscule AI startup raises $41 million to tap India growth - Times of India

Indian AI startup Sarvam AI raises $41 million in funding round, largest for an early-stage AI company in India

Sarvam AI aims to build affordable language models for unique uses in Indian languages [ more ]

www.cnbc.com

Meta's AI chief doesn't think AI super intelligence is coming anytime soon, and is skeptical on quantum computing

Yann LeCun believes current AI systems are decades away from reaching sentience and common sense capabilities.

LeCun believes the technology industry's current focus on language models and text data will not be enough to create advanced human-like AI systems. [ more ]

TNW | Deep-Tech

Google's Gemini AI won't be available in Europe - for now

Google has launched its new generative AI models called Gemini, which it claims to be the "most capable model ever."

Gemini models are trained to recognize, understand, and combine text, images, audio, video, and code. [ more ]

Theregister

Google unveils TPU v5p pods to accelerate AI training

Google has revealed its new performance-optimized chip, the TPU v5p, designed to reduce training time for large language models.

The TPU v5p is Google's most powerful chip yet, capable of pushing 459 teraFLOPS of performance and backed by 95GB of high bandwidth memory. [ more ]

WIRED

A New Trick Uses AI to Jailbreak AI Models-Including GPT-4

Large language models like ChatGPT have become popular among developers, with over 2 million using OpenAI's APIs.

These models can exhibit biases and fabricate information, leading to potential misuse and the need for safeguards. [ more ]

CodeProject

Fine-Tuning the Falcon 7-Billion Parameter Model with Hugging Face and oneAPI

Open-sourcing large language models makes AI technology more accessible.

Fine-tuning large language models involves adapting pretrained models for specific tasks. [ more ]

TNW | Deep-Tech

Silo AI releases checkpoint on mission to democratise LLMs

Large language models work more effectively in English, creating language bias and limiting access to knowledge and innovation in other languages.

Silo AI has released the multilingual open European LLM Poro 34B, which has shown best-in-class performance for low-resource languages like Finnish. [ more ]

time.com

AI and the Rise of Mediocrity

Artificial intelligence is not conscious or intelligent, but rather language and image models that predict patterns based on previous data.

AI tools are effective at regurgitating commonplace information, making lists, organizing notes, and generating basic content. [ more ]

www.vox.com

Why it's important to remember that AI isn't human

ChatGPT remains a topic of debate among experts, with opinions ranging from it being a potential threat to civilization to it being a sophisticated auto-complete tool.

The emergence of language models like ChatGPT raises questions about the link between language and the mind, and whether a new form of mind has been created.

Interacting with chatbots can be misleading due to the ambiguity in language, requiring us to rely on our intention-guessing mechanism for effective communication. [ more ]

Acm

What Would the Chatbot Say?

GPT-4, a large language model by OpenAI, demonstrated emergent behavior by creating a primitive image of a unicorn when asked to draw one

Emergent behavior is the ability of AI models to exhibit unexpected abilities beyond their training

Researchers are studying how these emergent abilities occur in language models [ more ]

SiliconANGLE

6 months ago

OpenAI reveals new details about its AI development roadmap and fundraising plans - SiliconANGLE

OpenAI is working on GPT-5 and plans to raise more capital from Microsoft to support its development.

GPT-5 will require more data to train than previous models and OpenAI plans to source that information from other organizations and publicly available sources.

Building large-scale language models is expensive and OpenAI intends to raise more funds to finance its development efforts. [ more ]

SG About Amazon

AI Singapore brings inclusive Generative AI models to Southeast Asia with AWS

Generative AI needs to become more culturally aware by training large language models on diverse data.

Organizations need to be able to customize their language models with local data in their native languages for social inclusion and economic growth. [ more ]

Open Data Science - Your News Source for AI, Machine Learning & more